Analysis framework of dependency measures in the stock market, and its applications

Abstract

Correlation coefficient has been the major dependency measure of detecting similar movements between stocks and financial markets in the field of financial research. However, due to the statistical assumptions whilst calculating the correlation, a more general dependency measures have been proposed: mutual information, Granger causality and transfer entropy. This study aims to compare the dependency measures by analysing how the measures imply the structural properties of the stock market, utilizing the graph theory. According to the analysis, we were able to find that relationship based measures like correlation and mutual information depicted graphs more closer to a random graph compared to the causality measures. Especially, during the financial crisis induced by COVID-19, we were able to detect that relationship based measures were not able to efficiently identify the structural properties. Using the properties of each dependency measures, this study constructed a portfolio based on the measures and analyzed the performances during the COVID-19 period. According to the performance results, the causality measures displayed better returns and a lower maximum drawdown. The correlation measure having the worst performances, we were able to conclude that causality measures are a better approach in explaining the stock market structure compared to the relationship measures.